5D Library › 算法营销、多触点沙普利归因与获客漏斗的随机控制 · English
第五章 一位客户还值多少钱:BG/NBD、Gamma-Gamma 与留存率的假象
非契约关系中的客户
订阅、电信与保险是契约关系:客户什么时候离开,账上有记录。电商、餐饮与直播带货多是非契约关系:客户最后一次购买之后再也没出现,我们不知道他是已经流失,还是只是还没到下一次购买的时候。客户终身价值(CLV)的估计在这两种情形下需要不同的模型 (Gupta et al. 2006)。本章先讨论非契约情形的 BG/NBD 与 Gamma-Gamma 模型,再讨论契约情形下「留存率随时间上升」的现象,以及它与辛普森悖论的区别。
BG/NBD 模型
施密特莱因(Schmittlein)、莫里森(Morrison)与科伦坡(Colombo)1987 年提出的 Pareto/NBD 模型是这一领域的起点 (Schmittlein et al. 1987),但它的似然函数需要反复计算高斯超几何函数,数值上相当麻烦。费德(Fader)、哈迪(Hardie)与李(Lee)2005 年提出了一个更容易计算的替代模型 BG/NBD (Fader et al. 2005a),假设如下:
- 客户在「活跃」期间按泊松过程购买,购买率为 ;
- 不同客户的 服从伽马分布 ;
- 每次购买之后,客户以概率 永久变为不活跃;
- 不同客户的 服从贝塔分布 ;
- 与 在客户之间相互独立。
与 Pareto/NBD 的区别在于第 3 条:流失只能发生在购买之后,而不是在连续时间中的任何时刻。这使似然函数只含伽马函数与贝塔函数。对一个在 内重复购买 次、最后一次在 的客户,
只依赖 三个数,即营销中常说的频率与最近一次购买(RFM 中的 F 与 R)。期望购买次数为
由似然的两项之比,可以得到一个客户在 时仍然活跃的后验概率: 时为 , 时为 1(模型中流失只发生在购买之后)。
证据
原文的数据与估计。 费德等人用 CDNOW 网站 1997 年第一季度首购客户的 1/10 系统抽样(2,357 人)、前 39 周为校准期、后 39 周为保留期,估计得 、、、,对数似然 −9,582.4;同一数据上 Pareto/NBD 的对数似然为 −9,595.0 (Fader et al. 2005a)。他们指出,对给定参数计算一次高斯超几何函数并不难,即使在电子表格中也可以用多项式级数近似。
计算 :级数与积分
高斯超几何函数 在 时收敛,但 越接近 1 收敛越慢。在 中 ,预测期越长越接近 1。另一种算法是欧拉积分表示 (Abramowitz and Stegun 1964)
这里有两个数值陷阱。第一,在 BG/NBD 中 ,被积函数在 处发散;第二, 接近 1 时, 在 附近急剧变大。程序的处理是:先用普法夫(Pfaff)变换 把自变量移到负半轴,此时 在 上有界;再令 ,把 变成 ,消去端点的奇异性,然后用辛普森法则积分。
推导
# BG/NBD (Fader, Hardie & Lee 2005). While alive, a customer buys as Poisson(lambda),
# lambda ~ Gamma(r, alpha); after each purchase she drops out with probability p,
# p ~ Beta(a, b). A: E[X(t)] = (a+b-1)/(a-1) [1 - (alpha/(alpha+t))^r 2F1(r, b; a+b-1;
# t/(alpha+t))]; 2F1 by its power series and by Euler's integral after Pfaff's
# transformation and a change of variable, compared as z -> 1, and checked against a
# simulation. B: a FICTIONAL cohort of 2,357 customers simulated with the published
# CDNOW estimates, refitted by maximum likelihood over 39 weeks; holdout forecast for
# weeks 40-78; P(alive) for two customers with the same frequency and recency apart.
import math, random, statistics
def f21_series(a, b, c, z, tol=1e-12, kmax=200_000):
term, s, k = 1.0, 1.0, 0
while abs(term) > tol * abs(s) and k < kmax:
term *= (a + k) * (b + k) / ((c + k) * (k + 1)) * z; s += term; k += 1
return s, k
def f21_euler(a, b, c, z, n=200):
"""For 0 < a < c, z < 1: Pfaff's transformation 2F1(a,b;c;z) = (1-z)^(-a)
2F1(a, c-b; c; w), w = z/(z-1) <= 0, then Euler's integral G(c)/(G(a)G(c-a))
int_0^1 s^(a-1) (1-s)^(c-a-1) (1-ws)^(b-c) ds, whose integrand stays bounded;
with s = u^(1/a) the weight s^(a-1) ds becomes du / a. Simpson's rule in u."""
w = z / (z - 1)
g = lambda u: ((1 - u ** (1 / a)) ** (c - a - 1)
* (1 - w * u ** (1 / a)) ** (b - c) / a)
h = 1 / n; tot = g(0) + g(1)
tot += sum((4 if i % 2 else 2) * g(i * h) for i in range(1, n))
lg = math.lgamma(c) - math.lgamma(a) - math.lgamma(c - a)
return (1 - z) ** (-a) * math.exp(lg) * tot * h / 3
def ex(t, r, al, a, b, f21=f21_euler):
z = t / (al + t)
return (a + b - 1) / (a - 1) * (1 - (al / (al + t)) ** r * f21(r, b, a + b - 1, z))
def simulate(rng, n, r, al, a, b, T):
out = []
for _ in range(n):
lam, p = rng.gammavariate(r, 1 / al), rng.betavariate(a, b)
t, times = 0.0, []
while True:
t += rng.expovariate(lam)
if t > T:
break
times.append(t)
if rng.random() < p:
break
out.append(times)
return out
def loglik(th, data):
r, al, a, b = th
if min(th) <= 0:
return -1e18
c = math.lgamma(r) - r * math.log(al)
lb = lambda x, y: math.lgamma(x) + math.lgamma(y) - math.lgamma(x + y)
ll = 0.0
for x, tx, T in data:
A1 = math.lgamma(r + x) - c
t1 = lb(a, b + x) - lb(a, b) + A1 - (r + x) * math.log(al + T)
if x > 0:
t2 = lb(a + 1, b + x - 1) - lb(a, b) + A1 - (r + x) * math.log(al + tx)
m = max(t1, t2); ll += m + math.log(math.exp(t1 - m) + math.exp(t2 - m))
else:
ll += t1
return ll
def p_alive(x, tx, T, r, al, a, b):
if x == 0:
return 1.0
return 1 / (1 + a / (b + x - 1) * ((al + T) / (al + tx)) ** (r + x))
def nelder_mead(f, x0, it=300):
n = len(x0)
pts = [list(x0)] + [[x * (1.2 if i == j else 1) for j, x in enumerate(x0)]
for i in range(n)]
for _ in range(it):
pts.sort(key=f); c = [sum(p[i] for p in pts[:-1]) / n for i in range(n)]
w = pts[-1]; r = [c[i] + (c[i] - w[i]) for i in range(n)]
if f(r) < f(pts[0]):
e = [c[i] + 2 * (c[i] - w[i]) for i in range(n)]
pts[-1] = e if f(e) < f(r) else r
elif f(r) < f(pts[-2]):
pts[-1] = r
else:
k = [c[i] + 0.5 * (w[i] - c[i]) for i in range(n)]
if f(k) < f(w):
pts[-1] = k
else:
pts = [pts[0]] + [[pts[0][i] + 0.5 * (p[i] - pts[0][i]) for i in range(n)]
for p in pts[1:]]
return min(pts, key=f)
if __name__ == "__main__":
PUB = (0.243, 4.414, 0.793, 2.426) # CDNOW estimates in FHL (2005)
r, al, a, b = PUB
print("A. 2F1(r, b; a+b-1; z) with the published parameters")
print(" t (weeks) z series (terms) Euler integral E[X(t)]")
for t in (39, 78, 520, 5200):
z = t / (al + t); s, k = f21_series(r, b, a + b - 1, z)
e = f21_euler(r, b, a + b - 1, z)
print(f" {t:5d} {z:.5f} {s:12.6f} ({k:6,d}) {e:12.6f}"
f" {ex(t, *PUB):6.3f}")
rng = random.Random(2005)
sim = simulate(rng, 20_000, *PUB, 78)
x39 = statistics.fmean(sum(x <= 39 for x in s) for s in sim)
x78 = statistics.fmean(len(s) for s in sim)
print(f" simulation, 20,000 customers: E[X(39)] {x39:.3f}, E[X(78)] {x78:.3f}")
print("B. fictional cohort of 2,357 simulated with the published parameters")
cohort = simulate(random.Random(1997), 2357, *PUB, 78)
cal = [(sum(t <= 39 for t in s), max([t for t in s if t <= 39], default=0.0), 39.0)
for s in cohort]
th = nelder_mead(lambda p: -loglik(p, cal), [0.5, 5.0, 0.5, 2.0])
print(" refitted r {:.3f}, alpha {:.3f}, a {:.3f}, b {:.3f}; log-likelihood {:.1f}"
.format(*th, loglik(th, cal)))
hold = sum(sum(39 < t <= 78 for t in s) for s in cohort)
pred = 2357 * (ex(78, *th) - ex(39, *th))
print(f" repeat purchases in weeks 40-78: actual {hold:,},"
f" BG/NBD forecast {pred:,.0f}")
for tx in (38.0, 10.0):
print(f" 4 purchases, last in week {tx:4.1f}: P(alive at week 39)"
f" {p_alive(4, tx, 39.0, *th):.3f}")
# A. 2F1(r, b; a+b-1; z) with the published parameters
# t (weeks) z series (terms) Euler integral E[X(t)]
# 39 0.89833 1.937370 ( 211) 1.937370 1.197
# 78 0.94644 2.390006 ( 401) 2.390005 1.861
# 520 0.99158 4.674606 ( 2,409) 4.674604 4.975
# 5200 0.99915 11.896629 (21,479) 11.896622 12.148
# simulation, 20,000 customers: E[X(39)] 1.174, E[X(78)] 1.830
# B. fictional cohort of 2,357 simulated with the published parameters
# refitted r 0.253, alpha 5.014, a 1.066, b 3.065; log-likelihood -10396.1
# repeat purchases in weeks 40-78: actual 1,446, BG/NBD forecast 1,329
# 4 purchases, last in week 38.0: P(alive at week 39) 0.838
# 4 purchases, last in week 10.0: P(alive at week 39) 0.055A 部分中,两种算法在 周时都给出 1.937370; 周(100 年)时 ,级数需要 21,479 项才收敛,积分只用 201 个节点,二者相差不到百万分之一(独立核查用 连接公式计算,表明此处级数值更准确)。、,2 万个模拟客户的平均值为 1.174 与 1.830,在抽样误差之内。
B 部分用原文参数模拟了一个虚构的 2,357 人队列,再用前 39 周的数据重新估计参数。 与 估得较准(0.253、5.014), 与 偏差较大(1.066、3.065),因为在只有 39 周的数据中,「流失」与「购买率低」难以区分。保留期(第 40–78 周)的实际重复购买为 1,446 次,模型预测 1,329 次,低估约 8%。最后两行说明 的用处:两个客户都买了 4 次,最后一次分别在第 38 周与第 10 周,前者仍活跃的概率为 0.838,后者只有 0.055。频率相同而最近一次购买不同,价值相差很大,这是「留存率相乘」类方法看不到的。
Gamma-Gamma 客单价模型
BG/NBD 预测购买次数,不预测金额。费德、哈迪与李在另一篇 2005 年的论文中把金额模型与购买模型结合起来,计算客户终身价值 (Fader et al. 2005b);所用的客单价模型的思想来自科伦坡与江(Jiang)1999 年 (Colombo and Jiang 1999)。设一位客户每笔交易的金额服从 , 在客户之间服从 ;给定这位客户 笔交易的平均金额 ,他未来每笔交易的期望金额为
即总体均值 与个人平均 的加权平均,交易次数越多,个人平均的权重越大 (Fader and Hardie 2013)。在 CDNOW 数据上(2,357 人中有重复购买的 946 人),估计值为 、、。
模型的一条关键假设是:客户之间平均交易金额的分布与购买过程相互独立 (Fader and Hardie 2013)。这样才能把「未来购买次数」与「每笔金额」分开估计再相乘。
推导
# Gamma-Gamma spend model (Colombo & Jiang 1999; Fader, Hardie & Lee 2005b): each
# transaction value ~ Gamma(p, nu), nu ~ Gamma(q, gamma) across customers, ASSUMED
# independent of the purchase process. A: fictional repeat buyers simulated with the
# published CDNOW estimates (p 6.25, q 3.74, gamma 15.44), refitted by maximum
# likelihood. B: expected spend E(Z | x, zbar) and a 104-week discounted CLV from the
# BG/NBD conditional expectation (FHL 2005a, eq. 10) for three fictional customers.
# C: when heavy buyers spend less per order, multiplying expected transactions by
# expected spend (the independence assumption) overstates revenue.
import math, random, statistics
# --- helpers repeated from earlier programs, so that this one runs on its own ---
def f21_series(a, b, c, z, tol=1e-12, kmax=200_000):
term, s, k = 1.0, 1.0, 0
while abs(term) > tol * abs(s) and k < kmax:
term *= (a + k) * (b + k) / ((c + k) * (k + 1)) * z; s += term; k += 1
return s, k
def nelder_mead(f, x0, it=300):
n = len(x0)
pts = [list(x0)] + [[x * (1.2 if i == j else 1) for j, x in enumerate(x0)]
for i in range(n)]
for _ in range(it):
pts.sort(key=f); c = [sum(p[i] for p in pts[:-1]) / n for i in range(n)]
w = pts[-1]; r = [c[i] + (c[i] - w[i]) for i in range(n)]
if f(r) < f(pts[0]):
e = [c[i] + 2 * (c[i] - w[i]) for i in range(n)]
pts[-1] = e if f(e) < f(r) else r
elif f(r) < f(pts[-2]):
pts[-1] = r
else:
k = [c[i] + 0.5 * (w[i] - c[i]) for i in range(n)]
if f(k) < f(w):
pts[-1] = k
else:
pts = [pts[0]] + [[pts[0][i] + 0.5 * (p[i] - pts[0][i]) for i in range(n)]
for p in pts[1:]]
return min(pts, key=f)
# --- end of repeated helpers ---
BG = (0.243, 4.414, 0.793, 2.426)
def gg_ll(th, data):
p, q, g = th
if min(th) <= 0:
return -1e18
ll = 0.0
for x, zbar in data:
ll += (math.lgamma(p * x + q) - math.lgamma(p * x) - math.lgamma(q)
+ q * math.log(g) + (p * x - 1) * math.log(zbar) + p * x * math.log(x)
- (p * x + q) * math.log(g + x * zbar))
return ll
def e_spend(x, zbar, p, q, g):
return p * (g + x * zbar) / (p * x + q - 1)
def ey(t, x, tx, T, r, al, a, b):
z = t / (al + T + t)
num = (a + b + x - 1) / (a - 1) * (1 - ((al + T) / (al + T + t)) ** (r + x)
* f21_series(r + x, b + x, a + b + x - 1, z)[0])
den = 1 + (x > 0) * a / (b + x - 1) * ((al + T) / (al + tx)) ** (r + x)
return num / den
def clv(x, tx, T, m, weeks=104, annual=0.10):
d = (1 + annual) ** (1 / 52); prev, tot = 0.0, 0.0
for w in range(1, weeks + 1):
cur = ey(w, x, tx, T, *BG); tot += (cur - prev) / d ** w; prev = cur
return tot * m, prev
def spenders(rng, n, p, q, g, dep=0.0):
out = []
for _ in range(n):
x = 1 + min(int(rng.expovariate(0.4)), 20) # repeat transactions
nu = rng.gammavariate(q, 1 / g) * math.exp(dep * (x - 3))
out.append((x, statistics.fmean(rng.gammavariate(p, 1 / nu) for _ in range(x)),
p / nu)) # (x, zbar, true mean spend)
return out
def pearson(x, y):
"""Pearson correlation (statistics.correlation needs Python 3.10+)."""
mx, my = statistics.fmean(x), statistics.fmean(y)
sxy = math.fsum((a - mx) * (b - my) for a, b in zip(x, y))
sxx = math.fsum((a - mx) ** 2 for a in x); syy = math.fsum((b - my) ** 2 for b in y)
return sxy / math.sqrt(sxx * syy)
if __name__ == "__main__":
PUB = (6.25, 3.74, 15.44)
data = spenders(random.Random(1999), 946, *PUB)
th = nelder_mead(lambda t: -gg_ll(t, [d[:2] for d in data]), [3.0, 3.0, 10.0], 400)
mean_pop = th[0] * th[2] / (th[1] - 1)
print(f"A. 946 fictional repeat buyers: refitted p {th[0]:.2f}, q {th[1]:.2f},"
f" gamma {th[2]:.2f}")
print(f" population mean spend {mean_pop:.2f}"
f" (published parameters: {PUB[0] * PUB[2] / (PUB[1] - 1):.2f})")
print(f" correlation of x with average spend: "
f"{pearson([d[0] for d in data], [d[1] for d in data]):+.3f}")
print("B. 104-week CLV, 10 % a year, published BG/NBD and Gamma-Gamma parameters")
print(" customer (x, t_x, zbar) E[spend] E[orders, 104 wk] CLV")
for x, tx, zb in ((1, 30.0, 60.0), (6, 38.0, 25.0), (6, 12.0, 25.0)):
m = e_spend(x, zb, *PUB); v, n = clv(x, tx, 39.0, m)
print(f" ({x}, {tx:4.1f}, {zb:4.1f}) {m:6.2f} {n:6.2f}"
f" {v:7.2f}")
print("C. 1,500 fictional buyers: revenue per buyer E[x m] (x repeat orders, m mean")
print(" spend per order) against E[x] E[m], the product used when spend is")
print(" modelled separately from frequency")
for dep, lab in ((0.0, "independent"), (0.25, "heavy buyers spend less")):
d2 = spenders(random.Random(7), 1500, *PUB, dep)
xm = statistics.fmean(d[0] * d[2] for d in d2)
pm = statistics.fmean(d[0] for d in d2) * statistics.fmean(d[2] for d in d2)
print(f" {lab:24s} E[x m] {xm:6.1f} E[x] E[m] {pm:6.1f}"
f" ({pm / xm - 1:+.0%})")
# A. 946 fictional repeat buyers: refitted p 7.72, q 3.50, gamma 11.88
# population mean spend 36.60 (published parameters: 35.22)
# correlation of x with average spend: -0.075
# B. 104-week CLV, 10 % a year, published BG/NBD and Gamma-Gamma parameters
# customer (x, t_x, zbar) E[spend] E[orders, 104 wk] CLV
# (1, 30.0, 60.0) 52.45 1.48 71.41
# (6, 38.0, 25.0) 25.70 8.40 198.96
# (6, 12.0, 25.0) 25.70 0.20 4.73
# C. 1,500 fictional buyers: revenue per buyer E[x m] (x repeat orders, m mean
# spend per order) against E[x] E[m], the product used when spend is
# modelled separately from frequency
# independent E[x m] 106.6 E[x] E[m] 109.4 (+3%)
# heavy buyers spend less E[x m] 85.3 E[x] E[m] 126.7 (+49%)A 部分用原文参数模拟 946 个虚构的重复购买者,重新估计得 、、;三个参数单独看差别不小,但它们决定的总体平均客单价相近(36.60 对原文参数的 35.22),这是这类模型中参数弱可识别的常见情形。B 部分结合 BG/NBD 的条件期望 与 Gamma-Gamma 的期望金额,按年 10% 贴现计算三个虚构客户未来 104 周的价值:只有一次重复购买、但那一次金额很高(60)的客户,期望客单价被收缩到 52.45,价值 71.41;买过 6 次、最后一次在第 38 周的客户价值 198.96;同样买过 6 次但最后一次在第 12 周的客户,期望只剩 0.20 次购买,价值 4.73。
C 部分检验独立性假设。如果常买的客户每笔花得少(作者判断这在促销驱动的电商中可能出现,见本章末的假说),收入的真实期望 为 85.3,而把频率与金额分开估计再相乘得到的 为 126.7,高估 49%;独立时二者只差 3%。所以在使用 Gamma-Gamma 之前,应当检查交易次数与平均金额的相关性(原文的 CDNOW 数据中二者相关很弱);相关明显时,需要联合建模。
「留存率随时间上升」:异质性,而不是辛普森悖论
契约业务中常见这样一张表:第一年续约率 63%,第二年 74%,第三年 82%……看起来客户「越来越忠诚」。作者框架把它称为辛普森悖论造成的假象。更准确的说法是:这是异质性造成的分选效应。费德与哈迪 2007 年用移位贝塔几何(sBG)模型说明了这一点 (Fader and Hardie 2007):每个客户每期以自己的概率 流失, 在客户之间服从 ;个人的流失概率不随时间变化。第 期的留存率为
随 单调上升,只是因为流失倾向高的人先走了。他们写道,在这个模型中个人层面没有任何时间动态,观察到的留存率上升「只是异质人群中的分选效应」。沃佩尔(Vaupel)与亚辛(Yashin)在人口学中系统讨论过同样的现象:群体的死亡率曲线可以与每一个个体的死亡率曲线形状完全不同 (Vaupel and Yashin 1985)。
推导
# Rising retention from heterogeneity, and why multiplying retention rates misleads.
# A: the shifted-beta-geometric (sBG) model of Fader & Hardie (2007): each customer
# churns each period with her own probability theta ~ Beta(alpha, beta); fitted by
# maximum likelihood to the published survival of their "regular" segment, years
# 1-7. Implied retention r_t = (beta + t - 1)/(alpha + beta + t - 1) rises with t
# although no individual's churn probability changes.
# B: projected survival and discounted expected lifetime (10 % a year) under the sBG
# and under the naive "year-1 retention multiplied" rule.
# C: a different phenomenon, a true Simpson's paradox: retention falls in both
# acquisition channels, yet the pooled rate rises because the channel mix shifted.
import math
SURV = [1.0, 0.631, 0.468, 0.382, 0.326, 0.289, 0.262, 0.241] # FH 2007, Table 1
def sbg_surv(t, al, be):
return math.exp(math.lgamma(be + t) - math.lgamma(be) + math.lgamma(al + be)
- math.lgamma(al + be + t))
def ll(al, be, n=1000):
if al <= 0 or be <= 0:
return -1e18
s = [sbg_surv(t, al, be) for t in range(8)]
lost = [n * (SURV[t - 1] - SURV[t]) for t in range(1, 8)]
return (sum(k * math.log(s[t] - s[t + 1]) for t, k in enumerate(lost))
+ n * SURV[7] * math.log(s[7]))
def fit():
best = (-1e18, 0, 0)
for i in range(1, 300): # coarse grid, then refine
for j in range(1, 300):
v = ll(i / 100, j / 100)
best = max(best, (v, i / 100, j / 100))
_, al, be = best; step = 0.005
while step > 1e-6:
moved = False
for da, db in ((step, 0), (-step, 0), (0, step), (0, -step)):
if ll(al + da, be + db) > ll(al, be):
al, be, moved = al + da, be + db, True
if not moved:
step /= 2
return al, be
if __name__ == "__main__":
al, be = fit()
print(f"A. sBG fitted to years 1-7: alpha {al:.3f}, beta {be:.3f}"
f" (paper: 0.704, 1.182)")
print(" year observed survival sBG survival sBG retention r_t")
for t in range(1, 8):
r = (be + t - 1) / (al + be + t - 1)
print(f" {t:4d} {SURV[t]:6.1%} {sbg_surv(t, al, be):6.1%}"
f" {r:6.1%}")
r1 = SURV[1]
print("B. projections from the same first-year retention of 63.1 %")
for t in (8, 12):
print(f" survival at year {t:2d}: sBG {sbg_surv(t, al, be):6.1%},"
f" naive 0.631^{t} = {r1 ** t:6.1%}")
d = 0.10
life_s = sum(sbg_surv(t, al, be) / (1 + d) ** t for t in range(0, 400))
life_n = sum(r1 ** t / (1 + d) ** t for t in range(0, 400))
print(f" discounted expected lifetime (years, incl. year 0, 10 %):"
f" sBG {life_s:.2f},"
f" naive {life_n:.2f} ({life_s / life_n - 1:+.0%})")
print("C. Simpson's paradox: 12-month retention by acquisition channel")
cohorts = {"2025": {"referral": (800, 0.60), "paid social": (3200, 0.20)},
"2026": {"referral": (3000, 0.55), "paid social": (1000, 0.15)}}
for y, ch in cohorts.items():
tot = sum(n for n, _ in ch.values()); kept = sum(n * p for n, p in ch.values())
print(f" cohort {y}: " + ", ".join(f"{k} {p:.0%} of {n:,}"
for k, (n, p) in ch.items())
+ f"; pooled {kept / tot:.1%}")
# A. sBG fitted to years 1-7: alpha 0.704, beta 1.182 (paper: 0.704, 1.182)
# year observed survival sBG survival sBG retention r_t
# 1 63.1% 62.7% 62.7%
# 2 46.8% 47.4% 75.6%
# 3 38.2% 38.8% 81.9%
# 4 32.6% 33.2% 85.6%
# 5 28.9% 29.2% 88.0%
# 6 26.2% 26.2% 89.8%
# 7 24.1% 23.9% 91.1%
# B. projections from the same first-year retention of 63.1 %
# survival at year 8: sBG 22.0%, naive 0.631^8 = 2.5%
# survival at year 12: sBG 17.0%, naive 0.631^12 = 0.4%
# discounted expected lifetime (years, incl. year 0, 10 %): sBG 3.69, naive 2.35 (+58%)
# C. Simpson's paradox: 12-month retention by acquisition channel
# cohort 2025: referral 60% of 800, paid social 20% of 3,200; pooled 28.0%
# cohort 2026: referral 55% of 3,000, paid social 15% of 1,000; pooled 45.0%A 部分用原文表 1 中「普通」客户前 7 年的存活比例拟合 sBG,得到 、,与原文完全一致;隐含留存率从第一年的 62.7% 升到第七年的 91.1%。B 部分比较两种外推:如果用第一年的留存率 63.1% 连乘,第 8 年只剩 2.5% 的客户,第 12 年 0.4%;sBG 给出 22.0% 与 17.0%。这两个数可以做样本外核对:Fader 等 2018 年公布了同一数据集后续年份的实际存活,第 8–12 年为 22.3%、20.7%、19.4%、18.3%、17.3% (Fader et al. 2018),sBG 只用前 7 年拟合,预测与之相当接近。按年 10% 贴现的期望存续年数,sBG 为 3.69 年,连乘只有 2.35 年;前者比后者高 58%,换言之连乘低估了约 36%。作者框架说「简单留存率相乘」估算 LTV 有误,这一点成立,而且方向通常是低估;费德与哈迪 2010 年专门讨论了在契约情形下忽略异质性给客户群估值带来的危害 (Fader and Hardie 2010)。图 4 画出了这两条曲线。
那么辛普森悖论是什么?它指的是:在每一个子群中,两个变量的关系方向相同,把子群合并后方向却反转 (Simpson 1951)。C 部分构造了一个营销中的例子:2026 年的队列中,推荐渠道的 12 个月留存率从 60% 降到 55%,付费社交渠道从 20% 降到 15%,两个渠道都变差了;但合并后的留存率却从 28.0% 升到 45.0%,因为 2026 年的新客户中推荐渠道的比例从 20% 升到了 75%。两种现象的共同点是「组成变化」;区别是:sBG 中没有任何子群的留存率在变化,变化的是存活者的组成;辛普森悖论中每个子群都在变差,变化的是新进入者的组成。前者的对策是用异质性模型外推,后者的对策是按渠道分层报告。
本章小结
BG/NBD 用四个参数描述非契约客户的购买与流失,只依赖频率与最近一次购买;它的期望公式中的 在长期预测时级数收敛很慢,用普法夫变换加变量代换后的欧拉积分可以稳定计算,两种算法互相核对。Gamma-Gamma 模型给出收缩后的期望客单价,但假设金额与频率独立;在常客花得少的虚构例子中,违反这一假设使收入高估 49%。逐期留存率随时间上升,通常是异质性造成的分选效应:sBG 精确重现了费德与哈迪的参数,并显示「留存率连乘」把期望存续年数低估了约三分之一。辛普森悖论是另一件事:子群都变差而合并后变好。
假说 · 尚未验证 · 直播电商客户的金额与频率负相关
作者的假说(尚未确立): 在以限时折扣为主的直播电商中,客户的重复购买次数与其平均订单金额负相关,使 Gamma-Gamma 与 BG/NBD 分开估计再相乘的客户价值系统性高于联合模型的估计。什么样的结果能证明它是错的: 在事先登记的、由独立研究者取得的若干个直播电商客户队列中,重复购买次数与平均订单金额的等级相关系数的合并估计,其 95% 置信区间包含零或为正。
参考文献
Fan, Y. (2026). Algorithmic Marketing, Multi-Touch Shapley Attribution and Stochastic Control of Acquisition Funnels: From Thompson Sampling Bandits and Absorbing Markov Chains to BG/NBD Customer Lifetime Value, Read Beside the Strategic Dispositions of the Sunzi [算法营销、多触点沙普利归因与获客漏斗的随机控制] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23836264 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).